The September work on Interview AiBox follows one question: does the assistant preserve what you prepared and what you actually said when the interview moves on? The most useful changes are often small: a clearer material selection, a correction that survives a long conversation, or a review that does not mistake an AI suggestion for your answer.

This article explains changes recorded in the repository between September 14 and 22, 2026. Implementation and local checks do not establish availability on every installed client or production deployment. Use the current release notes and the relevant feature guide to confirm your workflow. We are not announcing a new numbered desktop release here.

Select and read materials without guessing the click target

Recent planning changes align the selection and preview layout for personal, job, and supporting materials. The supporting-material list uses explicit Add and Added actions. Clicking a row opens the content; the action on the right determines whether it joins the current selection. A wider selection pane and overflow constraints help keep that action readable.

The trade-off is a quieter interface with fewer duplicated controls. The preview and the primary completion action have distinct jobs, and capacity is shown without repeatedly interrupting reading. That does not change plan limits or make a material ready before processing finishes.

For example, before an infrastructure interview, inspect a migration note before adding it. Seeing the note in preview is not the same as associating it with the interview. Confirm the saved plan and the titles selected on Desktop. The planning guide explains how materials reach the session.

Keep the current question and the latest correction available

A useful question can already be in the final model input while a coverage check incorrectly treats its source turn as missing. The latest Voice Ask changes record which real turn owns the rendered current question and let the existing coverage check recognize that evidence. This addresses a bookkeeping failure; it does not guarantee that every generated answer will be correct.

The update also preserves the latest accepted low-confidence turn until a validated context projection replaces it. A low speech-recognition confidence value is a reason to be cautious, not enough by itself to erase a potentially important correction.

Imagine a candidate saying, “I helped with the migration,” then clarifying, “I owned the rollback script, not the overall design.” Losing the clarification can produce a much more serious mistake than retaining a little extra transcript. The system therefore keeps raw evidence available when asynchronous cleanup cannot be verified. The cost is that some extra context remains in the request.

Improve material generation recovery without inventing facts

Candidate-profile generation needs a complete, parseable result. The recent fallback change uses a bounded JSON response mode for a specific provider path that could otherwise produce an unsuitable tool-call envelope. Users should care about the outcome: fewer format-related dead ends in that path, while the same validation and editable material boundary remain.

This is a targeted compatibility change, not a claim that provider errors disappear. It also does not verify the truth of extracted facts. Review your role, dates, responsibilities, and results before confirming materials in the Interview Materials workflow.

We chose a constrained fallback over asking users to repair raw model output. We also kept background material processing separate from the live answer path: a complete editable profile and the first useful words of a live response have different success criteria. No new performance percentage or latency promise follows from this work.

Review what the candidate said, with evidence limits intact

Recent review-sync changes avoid creating a review row from an empty or AI-only snapshot merely because client metadata suggests one. Empty transcript rows are omitted from displayed output. Session retention, visibility, and eligibility for a scored report are separate checks; this change is not a promise that every screenshot-only session disappears from every history view.

The broader product rule remains that AI suggestions are not candidate performance evidence. An impressive generated response cannot prove you said it. A session without enough real interviewer-and-candidate dialogue should not become a confident performance assessment.

Use the review guide to inspect the real exchange before using a score. Transcription alone cannot establish tone, accent, or emotional state. The deliberate trade-off is less apparent completeness: showing insufficient evidence is more useful than manufacturing a finished-looking report.

What we chose to prioritize

We prioritized usable evidence over more answer categories, more client-side decisions, or longer prompts. Material selection should show what you chose. Voice context should preserve the question and corrections. Derived summaries should remain accountable to their sources. Reviews should distinguish recorded speech from AI assistance.

These are design priorities, not guarantees of perfect memory or factual accuracy. Conservative recovery can retain noise and increase context size; provider fallback can still fail; a cleaner review list does not by itself establish better scoring. We have not derived a production accuracy gain or net cost saving from the local regression checks.

For your next rehearsal, try one material-selection change, one explicit correction during a conversation, and one post-session evidence check. Compare what you selected, what you actually said, and what the assistant used. That small loop is the practical reason behind the update.